2026-10-18
October 18, 2026
If agents can run the status loop, the program leader who survives is the one who owns the decisions nobody can delegate.
Two arguments about the future of engineering management are running past each other. One says the middle layer is dissolving: companies are flattening, and AI now drafts the status updates, summarizes the meetings and maintains the trackers that used to justify a coordinator. The other says management is about to matter more, because someone has to stay accountable for what humans and agents ship together. Both are partly right, and the useful question is which parts of your job sit on which side of that line.
James Stanier, CTO at Nordhealth, put the first half bluntly in an April 2026 LeadDev piece: the coordination-centered manager role is shrinking, driven by flattening and by AI tools that handle status, tracking, summaries and scheduling. His conclusion is the interesting part. People management is necessary but not sufficient; the managers who stay valuable combine leadership with real technical engagement, while those who mainly relay information are the most exposed.
TPMs should read that as addressed to them, only more so. Relaying information across teams is the textbook description of the weak version of the job. If your value is that you know what is happening everywhere and tell people, a model with read access to the tracker does a decent imitation of you. If your value is that you decide what the program is, what it is not, who decides what, and when to stop, nothing in the current tooling does that for you. The rest of this issue is about how to tell which job you are doing.
The flattening debate is usually conducted at the level of headcount ratios, which is the wrong altitude for an individual leader. Adam Ferrari, writing on his Engineering Together Substack in August 2025, suggests flattening has real benefits (less distortion as strategy cascades down, a higher "builder ratio") but also an equilibrium point beyond which outcomes degrade; he suspects ratios near ten ICs per manager may already be close to it for some team types, and notes that with fewer managers, each one has to be stronger. Meanwhile LeadDev reported in May 2026 that demand for engineering managers is rising in the agentic coding era, because directing agents looks a lot like what tech leads and managers already do: setting context, breaking down tasks, calibrating how much to delegate. Those two observations fit together if you stop treating "management" as one thing.
Here is a more useful decomposition for your own calendar. Every hour of a senior TPM's or director's week falls into one of three kinds of work.
Relay is moving information between people and systems without changing it. Status roll-ups, meeting notes, "where are we on X," chasing updates, keeping a tracker current. This is where automation is already good, and Stanier's list of what AI now handles maps almost exactly onto it. Expect the price of relay to approach zero. That is not a threat to the work; it is a threat to anyone whose entire value proposition is relay.
Judgment is work where the output is a decision or a framing that did not exist before: which of three dependencies to attack first, whether a slipping milestone is a schedule problem or a scope problem, what the real decision is behind an argument between two directors, which risk to escalate and with what recommendation. Agents can draft options for you. They cannot know what your organization will tolerate, which of your stakeholders is bluffing, or what you promised the VP in a hallway in March.
Accountability is being the named human who answers when it goes wrong. Ferrari's counterargument to full flattening rests on exactly this: humans must remain accountable for software quality, so a hybrid human-and-agent model still needs real management. For a TPM, accountability shows up as owning the commitment: the date you gave, the scope you agreed, the escalation you did or did not make. It cannot be delegated to an agent, because an agent cannot be held to anything.
The strategic move is to run the sort honestly, then act on it. Take last week's calendar and tag every block as relay, judgment or accountability. Most senior people discover that relay is far larger than they believed, often forty to sixty percent, because it is the work that feels productive and generates visible artifacts. Then do three things. Automate relay aggressively and without sentiment, and be the first person in your org to do it, not the last. Protect judgment time like deep work, because it is the part that compounds. And make accountability explicit: write down which commitments you personally own, so that when agents take over more of the flow, the human owner of each outcome is still a named person.
There is an uncomfortable corollary for leaders of leaders. If relay shrinks, the number of people needed to run a given program shrinks with it, and the remaining roles get more demanding, not less. The LeadDev piece quotes an engineering manager who says he is managing more people while doing much less management per person, because an agent now handles routing work that used to take him thirty to sixty minutes. Notice what that implies: the per-person management that remains is the coaching, calibration and conflict work, which is precisely the work that cannot be automated and precisely the work that gets squeezed when span of control widens. Widening spans without raising the quality of the remaining work is how flattening goes wrong.
A caution on evidence. The LeadDev demand piece rests mostly on interviews with a handful of executives and one anonymous manager, and its conclusions are predictions, not measured trends. Treat the direction as plausible and the magnitudes as unknown. What you can do with uncertain forecasts is what you do with any: build the capability that pays off across scenarios. Judgment and technical depth pay off whether the middle layer shrinks a lot or a little. Pure relay pays off in none of them.
Try this week. Export your last two weeks of calendar and tag every block relay, judgment or accountability. Pick the single largest relay block and automate or delete it by Friday. Then book two protected 90-minute blocks next week for judgment work, with a named decision as the output of each.
What it is. A way to hold three time frames in one view: Horizon 1 is the business and practices that run today, Horizon 3 is the emerging future that is still marginal, and Horizon 2 is the transitional work that bridges them. The McKinsey authors used it for portfolios of growth initiatives; Bill Sharpe's later version applies it to cultural and organizational change.
When to use it. When a team or program has to evolve its operating model while still delivering, such as moving a TPM function from relay-heavy to judgment-heavy, or adopting agent workflows without breaking current commitments.
How to run it:
When NOT to use it. In a crisis or a well-understood, stable process where there is no real choice to make about the future.
Example: H1 is the Monday status meeting, H3 is a live agent-generated program view, and H2 is one program running both for six weeks with the meeting shortened to decisions only.
The end of the non-technical engineering manager — James Stanier (LeadDev, April 2026) argues coordination-only management is the exposed part of the role; read it as a job description for the TPM who wants to stay valuable.
Demand for engineering managers is surging in the agentic coding era — Sage Lazzaro (LeadDev, May 2026) on why delegation and context-setting skills transfer to managing agents, with the caveat that it leans on a few interviews.
Will the Great Flattening eliminate engineering management? — Adam Ferrari's 2025 case that flattening has an equilibrium point, which is a useful frame when someone proposes widening your spans.
"Only the paranoid survive."
— Andy Grove, title of his 1996 book
Don't miss what's next. Subscribe to Critical Path: